Jing Li

PolMeth2024, APSA2025.

Both partisan polarization and partisan segregation are phenomena of great concern in the US in recent decades. Existing research that looks at mass partisan polarization and geography in conjunction focuses exclusively on the extent of partisan geographical sorting but neglects the substantial effect geographical context can have on the formation of mass partisan polarization. This paper shows that while partisan geographical sorting can happen to a certain extent, it only happens in spaces where there is a certain level of partisan segregation to start with. In essence, geographical context’s effect on partisan polarization is more substantial than partisanship’ effect on geographical sorting, a surprising result given current literature’s focus on the latter rather than the former.

Jing Li

PLOS ONE, 2024

The proper use of model evaluation metrics is important for model evaluation and model selection in binary classification tasks. This study investigates how consistent different metrics are at evaluating models across data of different prevalence while the relationships between different variables and the sample size are kept constant. Analyzing 156 data scenarios, 18 model evaluation metrics and five commonly used machine learning models as well as a naive random guess model, I find that evaluation metrics that are less influenced by prevalence offer more consistent evaluation of individual models and more consistent ranking of a set of models. In particular, Area Under the ROC Curve (AUC) which takes all decision thresholds into account when evaluating models has the smallest variance in evaluating individual models and smallest variance in ranking of a set of models. A close threshold analysis using all possible thresholds for all metrics further supports the hypothesis that considering all decision thresholds helps reduce the variance in model evaluation with respect to prevalence change in data. The results have significant implications for model evaluation and model selection in binary classification tasks.

Jing Li

Proceedings of 14th International Conference on Complex Networks and Their Applications (peer-reviewed and accepted), 2025

Existing research that studies the relationship between nationalized politics and partisan polarization focuses mostly on electoral politics, this paper examines the relationships between nationalization in terms of partisan interactions and mass partisan ideological polarization. In particular, through computational modeling, we highlight that as political communication becomes increasingly nationalized, individuals become more responsive to nationalized salient partisan news and media contents, which in turn leads to more intense partisan polarization process.

Jing Li

The synthetic control method is increasingly being used to estimate the effect of an intervention when units of interest are at the aggregate level and there does not exist suitable control units to serve as the counterfactual for the treated unit. However, the validity of the method has not been examined under complex settings where there could either be interaction effects among the covariates used to estimate the optimal synthetic control unit or where there could be a network cascading process happening within the treated unit at the same time as the intervention of interest. This paper shows that a number of commonly used synthetic control methods provide biased estimates of the causal effect of interest when there is only a small amount of pairwise interactions among the covariates and their coverage rates significantly decrease when there is an interaction between within treatment unit network cascading and the intervention of interest. As a result, this paper provides new insights about the validity of synthetic control methods under more complex analytical settings.

Jing Li

ACIC 2024

This paper examines the performance of bootstrap methods for quantifying the uncertainty of causal effect estimates from matching. In particular, it investigates the performance of three existing approaches as well as a newly one. The proposed non-parametric bootstrap method quantifies the uncertainty of average treatment effect estimate by first quantifying the uncertainty associated with the sample treatment group by bootstrapping the treatment group first and then finding the counterpart control group by pair matching on estimated propensity score without replacement. Through Monte Carlo simulation and analysis of the CPS data set, the results show that all bootstrap standard error estimates can deviate from the true sample standard error if the data becomes in-permissible for matching. And most bootstrap methods provide satisfactory coverage rate of 95 percent or above but only when percent treated in the sample data is small.